Research On Noise Detection Algorithm In Network Embedding

In the area of interconnection and big data,network structures have become the main form of storing data in fields such as: social network,recommendation system,and knowledge mapping.Network embedding,network representation algorithm,aims to present network by low-dimensional vectors,and preserves the network topology and rich auxiliary information in network.However,existing research on network embedding is based on cleaned data,which is almost non-existent in practical applications.Noise will bring negative impact both in network embedding and it's downstream tasks such as classification and link prediction.For this reason,it is necessary to study noise detection in network embedding.This thesis divides noise in network into two categories: label noise and edge noise,designing and implementing a noise detection algorithm framework in network embedding.The contents include research on label and edge noise detection,and research on high-noise rate issue,which cover different aspects of noise detection in network embedding.The main contents are described as follows:1.Label noise,an iterative semi-supervised learning detection algorithm is proposed.Based on research of integrated detection with multiple classifiers,the algorithm tackles the problem of less labeled data and difficulty in training classifiers by leveraging the adjacency relationship in network.The iterative framework improves noise detection and optimizes vertex's embedding.Experiments show that the algorithm is able to improve the accuracy of vertex's embedding on classification tasks.2.Edge noise,an unsupervised detection algorithm based on similarity is proposed.The algorithm divides noise into two categories: missing link and unnecessary link.Based on the link prediction to resolve missing link,the edge noise issue is solved.Experiments show that algorithm has higher accuracy in edge noise detection by detecting noise in different type.3.High noise rate,an integration model is proposed based on network structure: label integration and embedding integration.Based on statistical learning theory,the weight of crowdsourcing is estimated,which improves the effect of ensemble learning.Experiments prove the integration strategies by uniting correlation in network data is effective.